People should probably think about more ways to modify BPE tokenisers. I made this graphical representation of all the algorithms we have for this in the current literature (included in the ReBPE paper, aclanthology.org/2026.finding...)

BPE-knockout just got outperformed by an algorithm that modifies BPE tokenisers in a feedback loop to make them absorb more and more constraints. It doesn't even need more data to do that. It uses the tokeniser itself as a dataset. 🧵

The gadfly of {my immediate surroundings} has landed on Bluesky.